Tool index¶
Occam exposes 18 core tools under OCCAM_PROFILE=full (product default is reader = 11);
more are opt-in via host environment flags (batch/watch/consensus/atlas/browser/exam).
All tools return a JSON string (camelCase). The trust rule everywhere: ok: false means the
page content is unknown — never substitute model memory.
Per-tool pages below were rewritten from code (Status verified @ 2026-09-29) during the doc
rescue. Runtime tools/list remains authoritative for availability and input schemas.
For an agent, this page is a router, not a prompt payload: choose one tool, then open only its linked page.
Pick by job¶
| You want to… | Use | Notes |
|---|---|---|
| Tell Occam your context window (once) | occam_client_capabilities |
Sizes later reads; or set OCCAM_CLIENT_CONTEXT_TOKENS |
| Read one page (cascade: url / task / budget) | occam |
Progressive facade; default on minimal/basic |
| Read one page with full opt-ins | occam_transcode |
Power reader; only url required |
| Decide whether a page is worth fetching | occam_probe |
Cheap classify: extractability 0–1, recommended backend |
| Research several pages at once | occam_digest |
Up to 8 URLs → per-page excerpts + combined Markdown |
| Find pages on a site when you have no URLs | occam_map |
Same-domain links from homepage/sitemap/robots |
| Search the open web for URLs | occam_search |
Keyless DuckDuckGo by default; override or off via env |
| Extract typed fields (title/price/author…) | occam_extract_knowledge |
Needs a playbook knowledge_schema for the host |
| Look up a site's saved extraction recipe | occam_playbook_resolve |
Read-only |
| Draft a recipe for a hard site | occam_playbook_heal |
Captures DOM skeleton + selector candidates |
| Validate a recipe without fetching | occam_playbook_lint |
Static, deterministic |
| Save a recipe (with live verify) | occam_playbook_save |
Default dry-runs a transcode first |
| Verify a signed receipt / cite a block | occam_verify |
offline / live / prove / citation / history modes |
| Prove an agent read a page | occam_canary_issue → fetch → occam_canary_verify |
HMAC sentinel; four verdicts |
| Check whether a page backs a claim | occam_claim_check |
Returns retrieved blocks + Merkle membership proofs; found:false is retrieval-only |
| Audit a report's citations in bulk | occam_attest |
1–50 {claim, sourceUrl} rows |
| Build a signed, auditable URL corpus | occam_dataset_export |
Per-row receipts + one manifest signature |
Opt-in tools (absent from tools/list until enabled)¶
Set the flag in the host environment before starting the MCP server, then reload the client.
| Tool | Enable with | Purpose |
|---|---|---|
occam_batch_submit / occam_batch_status / occam_batch_results |
OCCAM_BATCH_MCP=1 |
Fire-and-forget async transcode of a URL list |
occam_watch |
OCCAM_WATCH_MCP=1 |
Stateful page-change detection; history entries signed when receipts enabled |
occam_crosscheck |
OCCAM_CONSENSUS_MCP=1 |
Compare vantage points; detect cloaking/personalization |
occam_failure_atlas |
OCCAM_ATLAS_MCP=1 |
Per-host failure map of the current run; skip walled hosts |
occam_browser_interact |
OCCAM_BROWSER_ACTIONS_MCP=1 |
Declarative browser actions, then materialize Markdown + Receipt v1 |
occam_exam_submit |
OCCAM_EXAM_MCP=1 |
Grade a capability-exam harness JSON; may change tools/list + list_changed |
occam_crawl |
OCCAM_CRAWL_MCP=1 |
Same-domain map→transcode crawl with resume_token + optional JSONL |
Common parameters¶
backend_policy—http(fast, no JS),browser(Playwright render), orhttp_then_browser(default: HTTP first, escalate when the result looks thin or challenged). See concepts.session_profile— id of a headers profile JSON underOCCAM_SESSIONS_ROOT/<id>.json, for gated/logged-in pages. See configuration.focus_query/fit_markdown/max_tokens— token-economy controls; see each page.
Related¶
- Failure codes — the full
failure.coderegistry - Receipts — signed extraction receipts
- Configuration — environment variables